Skip to main content
Vector search lets your chatbot find relevant documents from your knowledge base to answer questions accurately. How it works:
  1. User asks a question
  2. Question converted to vector embedding
  3. Similar vectors found in Pinecone
  4. Relevant documents retrieved
  5. Results reranked for relevance
  6. Documents used as context for response

​
The Flow

User Question
    ↓
Query Embedding
    ↓
Vector Search (Pinecone)
    ↓
Retrieve Documents
    ↓
Rerank Results
    ↓
Use as Context
    ↓
Generate Response

​
Configuration

​
Basic Setup

from langchat.vector_db import Pinecone

vector_db = Pinecone(
    api_key="pcsk-...",
    index_name="your-index",
    embedding_model="text-embedding-3-large"  # Optional
)

​
Retrieval Settings

Control how many documents to retrieve: 
# Access retriever directly
retriever = vector_db.get_retriever(k=10)  # Retrieve 10 documents
Recommendations:
  • Small index (< 1000 docs): k=5
  • Medium index (1000-10000 docs): k=10
  • Large index (> 10000 docs): k=20

​
Embedding Models

​
text-embedding-3-large (Recommended)

vector_db = Pinecone(
    api_key="...",
    index_name="...",
    embedding_model="text-embedding-3-large"
)
Advantages:
  • Highest quality
  • Best for semantic search
  • Good for complex queries

​
text-embedding-3-small

vector_db = Pinecone(
    api_key="...",
    index_name="...",
    embedding_model="text-embedding-3-small"
)
Advantages:
  • Faster and cheaper
  • Good for simple queries
  • Lower latency

​
Reranking

Reranking improves search result relevance: 
from langchat.reranker import Flashrank

reranker = Flashrank(
    model_name="ms-marco-MiniLM-L-12-v2",
    top_n=3  # Keep top 3 after reranking
)

ai = LangChat(
    llm=llm,
    vector_db=vector_db,
    db=db,
    reranker=reranker
)
Reranking models are automatically downloaded on first use.

​
Best Practices

​
1. Balance Retrieval Count

# Too small: Might miss relevant docs
k = 3

# Too large: Too many irrelevant docs
k = 50

# Optimal: Depends on your index
k = 10  # Good for most cases

​
2. Always Use Reranking

Reranking improves relevance: 
reranker = Flashrank(top_n=3)  # Rerank and keep top 3

​
3. Choose Right Embedding Model

# High quality (recommended)
embedding_model="text-embedding-3-large"

# Faster and cheaper
embedding_model="text-embedding-3-small"

​
Troubleshooting

​
No Relevant Results

Solutions:
  • Increase retrieval count: k=20
  • Check Pinecone index has documents
  • Try different embedding model

​
Too Many Irrelevant Results

Solutions:
  • Use reranking: top_n=3
  • Improve document quality
  • Use better embedding model
Solutions:
  • Reduce retrieval count: k=5
  • Use smaller embedding model
  • Optimize Pinecone index

​
Next Steps

Built with ❤️ by NeuroBrain